Inverse Materials Design: Joint Generation of Crystal Structures and Electronic Descriptors

npj Computational Materials

Overview

A new diffusion framework improves the inverse design of inorganic crystals by jointly denoising structural variables and local electronic descriptors. The model achieves higher success rates in property-conditioned generation while increasing the fraction of structures that are thermodynamically stable, unique, and novel. This approach mitigates the trade-off between target property accuracy and structural physical plausibility.

In Depth

Key Findings

The proposed joint diffusion framework outperforms structure-only baselines in generating inorganic crystals conditioned on band-gap, formation energy, or a combination of both. By incorporating local electronic descriptors, the model increases the proportion of generated structures that satisfy physical validity, novelty, uniqueness, and thermodynamic stability. Control experiments using dummy variables or noise-corrupted data indicate that these improvements depend on the joint variables faithfully reflecting the true local electronic structure.

Technical Details

The framework utilizes a shared score network to jointly denoise crystal-structure variables and site-resolved local electronic descriptors. The researchers adopted Bader charge and atomic density of states (atomic DOS) as the representative descriptors. Results show that generated Bader charges align with DFT references with a mean absolute error (MAE) of 5.4 × 10⁻² e for stable, unique, and novel structures. While accuracy varies by elemental species, the generated atomic DOS captures the coarse spectral profile of the DFT reference around the median of its accuracy distribution.

Background and Context

Inverse design—generating structures to satisfy specific target properties—is more challenging than ab initio generation. Current generative models often suffer from a degradation in structural quality or a collapse to a small subset of memorized structures when property conditioning is applied. While previous diffusion models like CDVAE, DiffCSP, DiffCSP++, SymmCD, and MatterGen have advanced crystal generation, the trade-off between property targeting and physical plausibility remains a central challenge.

Significance and Outlook

This research establishes that local electronic descriptors serve two complementary roles in materials discovery. First, they broaden the explored materials space by increasing structural diversity. Second, they guide the structural trajectory toward electronically plausible configurations during the joint denoising process. This methodology provides a more explicit description of chemical bonding and stability than structural variables alone, offering a more robust route for the practical inverse design of inorganic materials.

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

• Your email and selected fields are used only to deliver the newsletter.
• We never share your information with third parties.
• You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Agree & Continue ▶

Takes about a minute · Unsubscribe anytime